[ROCm] Auto-enable breakable CUDA graphs for DeepseekV41ForCausalLM (#56349)

Signed-off-by: Markus Hartikainen <markus.hartikainen@amd.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Markus Hartikainen
2026-09-12 20:47:07 -07:00
committed by GitHub
co-authored by Jee Jee Li
parent c2ea9f1a5f
commit c711f740b7
2 changed files with 14 additions and 3 deletions
+4
View File
@@ -358,6 +358,10 @@ def test_rocm_keeps_compiled_deepseek_defaults(monkeypatch):
assert "DeepseekV32ForCausalLM" not in breakable_architectures
assert "DeepseekV32MTPModel" not in breakable_architectures
assert "GlmMoeDsaForCausalLM" not in breakable_architectures
# V4.1 cannot torch.compile and the ROCm sparse SWA backend only
# supports uniform-batch CUDA graphs, so it must opt in or the
# default FULL_AND_PIECEWISE serve path cannot start.
assert "DeepseekV41ForCausalLM" in breakable_architectures
# The carve-out takes effect via the runner-selection property
# (warning_once args must be hashable for its lru_cache).